Survey Protocol Cards for Crop Maps

Akram Zaytar, Girmaw Abebe Tadesse, Caleb Robinson, Shabarinath Nair, Gerald Blasch, Jeroen Degerickx, Mitelo Subakanya, Juan Carlos Laso Bayas, Gilles Quentin Hacheme, Inbal Becker-Reshef, Rahul Dodhia, Juan Lavista Ferres

Crop type maps underpin food security decisions yet their accuracy depends on label quality, which in turn depends on survey design choices made under tight budgets. Survey planners must allocate limited resources across GPS devices, stratification strategies, sample size, worker training, and verification protocols, but lack quantitative guidance on which investments yield quality crop maps. We address this gap by modeling the full chain from survey design to downstream crop detection accuracy: survey choices map to costs, costs constrain achievable label noise levels, and noise levels affect crop mapping performance. We implement 17 noise functions grounded in documented errors from the agricultural survey literature, and measure degradation on two datasets: EuroCrops and Zambia. Our experiments reveal that label verification matters far more than GPS accuracy: crop misidentification causes up to 99% F1 loss while 30m GPS jitter causes only 4%. Dataset-specific noise-to-performance surrogate models achieve R2=0.87, enabling millisecond what-if queries---but cross-dataset transfer shows mixed results: Spearman ρ=0.32--0.60 indicates rankings transfer asymmetrically, and negative R2 reveals degradation predictions fail across contexts. We package these findings into a programmable protocol-card and web interface that optimizes survey design given budget constraints.